The post
hey all, my background is photography and ive only recently started messing with ai video. so far my setup has been pretty simple. i train a lora on pixel dojo, use that lora to make a base image, then feed the image into wan 2.6 to turn it into a clip. the trouble is how unpredictable it is. getting a start image i actually like can take a bunch of tries, and even with a good one the video step often needs several rerolls before anything looks usable. every retry costs something, so i can see the bill climbing fast before i end up with a finished video. im also fairly sure pixel dojo isnt the cheapest way to do this. one idea is to cut down the wasted rerolls by checking each base image with a cheap short test before i spend real credits on it. my own capcut account has an individual nearly-free offer for dreamina (campaign collaboration, just so youre aware), which makes one hosted test cost me very little.
so, my plan is to upload the lora base image and generate a short image-to-video clip from it. if the motion and framing hold up, that image moves on to the proper wan run. if it falls apart, i fix the image first instead of paying for more video attempts. the bigger question is the rest of the setup. ive been thinking about grabbing open source wan, managing things from my macbook air, and sending the image and video generation off to a cloud gpu service. has anyone here run it that way? did it actually save money once you counted setup time and gpu hours? and any other tips for keeping costs under control would really help. thanks
View the original thread on Reddit →